atq2119

55 ms·

541.
by atq2119 3y ago
> Has AI ever caused anything to go extinct? We know from human history that intelligence tends to cause extinctions. AI just hasn't been around long enough, nor been intelligent enough yet. Though, if you count corporations as ar
542.
by atq2119 3y ago
You're taking these things too lightly. It's true that most of these things are unlikely to kill all humans directly, but with most of them, civilizational collapse is definitely on the table, and that can ultimately lead to human
543.
by atq2119 3y ago
It's interesting that you call that area (from the TV show) a mall. It would never have occurred to me to call it that, it's just a normal mixed-use walkable part of a city.
544.
by atq2119 3y ago
You could look lower in the stack, e.g. at compilers, drivers, and libraries for GPUs and other accelerators. That's not a huge field, but still a decent number of jobs to go around.
545.
by atq2119 3y ago
In other words: you had it bad in the past, therefore the future is not allowed to be better. More seriously, while this particular discussion centers around US visas, obviously the analogous arguments apply to other countries as well.
546.
by atq2119 3y ago
The kernel/user space split in graphics drivers is pretty much the same between Linux and Windows, and has been this way for many, many years.
547.
by atq2119 3y ago
It's because the input tokens can be batch-processed in a single forward pass through the model, while generating tokens requires one forward pass through the model per token. If you do the math of how much memory bandwidth is required
548.
by atq2119 3y ago
Perhaps. But it was also still relatively early days for computers, and it was a time of union busting, macroeconomic upheaval, and a shift in dominant politics. There were certainly multiple causal factors at play. You say that there are t
549.
by atq2119 3y ago
The question is more about who gets to capture the created wealth: - people who gave money to enable the activity leading to profit (aka capital) - people who organize the activity leading to profit (aka management) - people who actually do
550.
by atq2119 3y ago
The parent comment is surely well aware of that. They were speaking normatively (what should be) rather than descriptively (what is).
551.
by atq2119 3y ago
If StackOverflow dies, surely the developer-relevant quality of training data will suffer? After all, the capabilities of ChatGPT are basically proportional to how well a topic is represented in the training data, which is largely the inter
552.
by atq2119 3y ago
Batch processing scales quadratically with the context size (assuming OpenAI is still using standard transformer architecture) but the batch processing of the prompt is also fast compared to generating tokens because it's batched (para
553.
by atq2119 3y ago
Right, makes sense. But then what do you actually do with a database? Starting with: what do you store in it? Maybe sentence/vector pairs. But what does that give you? What do you do with that data algorithmically? What's the eq
554.
by atq2119 3y ago
It should work. After the initial prompt processing, token generation is typically limited by memory bandwidth more than by compute.
555.
by atq2119 3y ago
It should be noted that that last part has a pretty large factor to it that also scales with model size, because to run transformers efficiently you cache some of the intermediate activations from the attention block. The factor is basicall
556.
by atq2119 3y ago
By this point, there can't be too many "Beamte" left in Germany's railway and public transit systems. After all, nobody there has received that status since the 90s.
557.
by atq2119 4y ago
The big advantage of the Eytzinger layout actually comes into play when your binary search isn't branch-free. In that case, the CPU will speculate loads reasonably far ahead, and because both choices tend to be in the same cacheline,
558.
by atq2119 4y ago
That blog slightly misrepresents futex. Or perhaps it's outdated. It's written as if futexes can only be used with a 1/0/negative protocol, but that's not true. The values used are up to user space.
559.
by atq2119 4y ago
The details are hazy for me but all relevant CPUs have coherent caches, but not all make the same ordering guarantees. x86 has "total store ordering", meaning stores made by core 1 will always be observed in-order by core 2. ARM d
560.
by atq2119 4y ago
True. Interestingly, while I'm pretty sure that I automatically do that correctly, I don't perceive it as a substantial difference. So perhaps part of the explanation is around how speakers of different languages / accents
561.
by atq2119 4y ago
There may well be the occasional service that drops 100x in price, but that would apply only to a small fraction of GDP. Most goods and services are already pretty well optimized. Amdahl's law applies and means that claims of rapid ord
562.
by atq2119 4y ago
Out of curiosity, does that mean you pronounce "historic" substantially different from "history"? And how about "the historic" as opposed to "a(n) historic"? I'm not a native speaker, but the tho
563.
by atq2119 4y ago
Of course it's about freedom. What some people don't want to hear is that their freedom must be limited where it impacts other people. Nobody is alone in the world.
564.
by atq2119 4y ago
I suspect the 1% number is typical only of programs that haven't been optimized. Once people start to care about performance and address these hotspots, which are often actually quite low hanging fruit, the profile starts to become fla
565.
by atq2119 4y ago
Take a look at Agner Fog's https://www.agner.org/optimize/instruction_tables.pdf For example, Zen4 64-bit DIV is listed as: 2 uOps, 10-18 cycles latency, 7-12 cycles inverse throughput. This suggests uOps with var
566.
by atq2119 4y ago
While division may still decode to multiple uOps, I seriously doubt that there's a loop in microcode on modern processors. The pipeline latency makes that infeasible. The looping logic is almost certainly a bit of fixed function hardwa
567.
by atq2119 4y ago
Sure, quite likely - but only with significant engineering effort. I said as much in my comment. The point is that it's not a turnkey solution today , but some of the comments on here make it sound like it is. That's a form of hy
568.
by atq2119 4y ago
> can take hundreds of pages of text and distill it down to an executive summary of any length This is the kind of breathless claim that no doubt fuels the skeptics. None of the context windows are large enough for "hundreds of page
569.
by atq2119 4y ago
One of the weaknesses of transformers is that inference is inherently serial, one token at a time, and the entire model needs to be read once per token. This means that inference (after prompt processing) is bounded by memory bandwidth inst
570.
by atq2119 4y ago
> On the subject of this thread, it was also DirectX 12 Ultimate that introduced Mesh Shaders. It was actually Nvidia who introduced mesh shaders by way of Vulkan and DX12 extensions. And then both Khronos and Microsoft went through conc

More ›